Enterprise Readiness for Agent AI Failures: A Survey Insight
As artificial intelligence becomes increasingly integral to business operations, the resilience of these systems is paramount. A recent survey conducted by Keepit, a Denmark-based cloud backup and recovery service, sheds light on enterprise preparedness for potential AI disruptions. This survey, encompassing over 300 IT decision makers from Australia, New Zealand, Europe, the United Kingdom, and the United States, reveals a significant gap between confidence and actual practice in disaster recovery.
Confidence vs. Practice in AI Disaster Recovery
While an impressive 94% of organizations expressed confidence that their disaster recovery plans adequately cover agentic AI systems, only 32% confirmed that these plans are tested monthly. This discrepancy raises questions about the true resilience of these enterprises in the face of AI-related disruptions.
Moreover, the survey highlighted a concerning reality: 33% of IT and security leaders admitted to having only partial control over the use of agentic AI within their organizations. Additionally, 52% of respondents expressed doubts about the comprehensiveness of their recovery plans for AI-specific scenarios.
Need for Structured and Frequent Testing
Kim Larsen, Group Chief Information Security Officer at Keepit, emphasized the importance of structured and regularly tested disaster recovery plans. “Organizations need to place more emphasis on creating long-term, structured and tested disaster recovery plans,” Larsen stated. He also highlighted the critical role of data governance and accountability in forming the foundation of any resilience plan.
Most organizations have evaluated large-scale data recovery at least once; however, consistent and systematic testing across all systems is lacking. A particularly neglected area is that of identity systems, such as Microsoft’s Entra ID and Confluence’s Okta, which are tested far less frequently than productivity applications like Microsoft 365, Google Workspace, and Salesforce.
Discrepancies in Recovery Efforts
The survey revealed that productivity applications are restored on average four times as often as identity applications. “Out of four organizations that conduct an annual test of their productivity workloads, only one of them (25%) has conducted a test of their identity applications,” the report noted. Most recovery activities focus on individual file downloads, reflecting routine operational needs rather than preparing for large-scale recovery events.
Learning from External Events
Keepit’s report aimed to assess whether significant external events influenced recovery behavior. Despite notable incidents such as the May 2024 solar flares, the July 2024 CrowdStrike incident, and the October 2025 Microsoft outages, there was no observed change in user behavior or increased backup activity post-events. This suggests either a lack of extensive remediation needs or a failure to leverage these events as catalysts for change in recovery routines.
To address this, the report’s authors suggest a proactive approach, using external events as triggers for guided recovery testing. Such tests should be short, repeatable validations that boost confidence without necessitating large-scale disruptions.
Implementing Guided Recovery
The report also advocates for “guided recovery” using MCP (Model Context Protocol). This mechanism enables administrators to request assistance at critical moments, identifying erroneous tenants or suspicious patterns in protected data and guiding them through correct recovery steps. This proactive strategy ensures a manageable and repeatable recovery process.
Larsen concluded by emphasizing the importance of clear responsibilities and priorities in recovery efforts. “It all boils down to knowing who is responsible for recovery and which systems will be restored first if multiple systems are affected. If decisions are delayed, recovery will take longer than necessary.”
The full report is available here on the Keepit website (registration required).
“`
